Speaker
Description
Consumer-grade wearables enables continuous monitoring and early detection of aberrant physiological signals associated with diverse health issues. However, substantial measurement noise and natural physiological variability in wearable data have prevented reliable identification of physiological anomalies, limiting their deployment in real-world and clinical settings. Here, we propose a real-time anomaly detection framework that first estimates latent physiological dynamics using a Kalman filter and subsequently detects anomalies using an autoencoder-based model. By explicitly modeling physiological dynamics, our approach projects out known sources of variation such as circadian rhythms, intrinsic biological fluctuations, and measurement noise, leaving anomalies visible in the residual. We tested our method on real-world wearable body temperature data from cancer patient and showed that it can detect early signals of impending fever events. Applying anomaly detection after physiological filtering outperforms an existing method of directly applying the autoencoder to raw data, including LSTM-autoencoder based models previously used for wearable anomaly detection. This indicates the importance of incorporating physiological structure and filtering into machine learning pipelines.
